RGBD Salient Object Detection via Disentangled Cross-modal Fusion.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Depth information enhances salient object detection (SOD) by providing additional cues.
• Existing RGBD SOD methods often use complex fusion strategies, risking overfitting and unclear cross-modal complementarity.

Purpose of the Study:

• To decouple diverse cross-modal complements for simplified and sufficient fusion in SOD.
• To reduce uncertainty and enhance adaptability in cross-modal fusion through explicit disentanglement.

Main Methods:

• Designed a disentangled cross-modal fusion network.
• Utilized cross-modal reconstruction to expose structural and content representations from RGB and depth data.
• Developed a fusion module to identify and incorporate desired complements based on disentangled representations.

Main Results:

• Demonstrated the effectiveness of the proposed disentangled fusion approach.
• Achieved significant outperformance over state-of-the-art SOD methods in extensive experiments.
• Showcased improved adaptability and reduced uncertainty in multi-modal fusion.

Conclusions:

• Explicitly disentangling cross-modal representations simplifies fusion and enhances its sufficiency for SOD.
• The proposed network offers a more robust and adaptable solution for RGBD salient object detection.
• This approach paves the way for more interpretable and effective multi-modal fusion techniques.